Dynamic House Allocations
Atila AbdulkadirogluColumbia University
Simon LoertscherUniversity of Bern∗
November 8, 2005
JOB MARKET PAPER
Abstract
We consider the problem of assigning individuals (e.g. students) toindivisible goods (schools) when these assignments have to be made re-peatedly and when individuals face uncertainty about the intensity of theirfuture preferences. For the setting with two schools and two periods, whereall individuals have an ordinal preference for the same school, we show thatexpected utility under a dynamic mechanism is strictly greater for everyindividual than under a static mechanism, where a mechanism is calleddynamic (static) if its future assignments depend (do not depend) on pastreports. We derive conditions under which a simple dynamic mechanismachieves the first-best allocation in the first period and second-best overboth periods. Extensions of the model allow, in turn, for individuals whodiffer with respect to their ordinal preferences, uncertainty about futureordinal preferences, correlated utilities, and the possibility that switchingschools is costly.
Keywords: House allocation, matching, mechanism design without trans-
fers, preference intensity.
JEL: C72, C78, D02
∗Abdulkadiroglu: Department of Economics, International Affairs Building, 420 West118th Street New York, NY 10027 Email: [email protected]. Loertscher: Departmentof Economics, Schanzeneckstrasse 1, CH-3001 Bern. Email: [email protected] acknowledge very valuable comments by Roland Hodler. The paper has also benefittedfrom comments of seminar participants at the University of Bern, Tilburg University (CentER)and at the Bolzano Summer School on Game Theory and its Applications 2005.
1
1 INTRODUCTION 2
1 Introduction
We consider the problem of assigning individuals (e.g. students) to indivisi-
ble goods (schools) when these assignments have to be made repeatedly and
when individuals face uncertainty about the intensity of their future prefer-
ences.1 These assumptions are satisfied in many real world situations, such as
the allocation of high school students to public schools, college students to cam-
pus houses, students to courses at business schools, or scarce resources within
an organization in general. Since for example students who are assigned to a
course in a given semester typically have to be assigned to a course the following
semester, it is clear that these problems are dynamic.2
Though this dynamic dimension adds naturally some complications, it also
offers potential for substantial efficiency gains. Very much like in the theory
of repeated games and the literature on dynamic contracting, future payoffs
can serve as incentives in the presence. For example, assume that there are
two schools and two periods and that all students agree which school is better.
If students who only weakly prefer the good school in period one are credibly
promised a sufficiently higher probability of getting into the good school in
period two when they apply to the bad school today, the allocation in period
one can eventually be improved. As those with the lowest preference for the
good school apply to the bad school today, there is either no rationing at the
good school in period one, or rationing is more efficient because it occurs among
students with a stronger preference for the good school.
Consequently, dynamic mechanisms have the potential of eliciting some car-
dinal utility information, namely whether individuals care a lot or a little about
the good school in period one. This contrasts with a static or myopic mecha-
nism, where in every period all students have a dominant strategy to apply to
their preferred school. Moreover, as a consequence of the increased efficiency
in the first period, ex ante expected utility under a dynamic mechanism (i.e.
expected utility of individuals before they know their valuations) exceeds ex
ante expected utility under a static mechanism.
The paper, and the model we adopt, are motivated by the problem of public
school choice, where a given number of students have to be assigned to a given1Allocation problems where indivisible goods - ”houses” - are assigned to individuals with-
out using monetary transfers have become known as house allocation problems.2Similarly, students who are assigned to a school in a given period will have to be assigned
to a school in the next period, and scarce resources within an organization are allocated tomembers of the organization repeatedly.
1 INTRODUCTION 3
number of schools.3 Because of either political reasons, credit market imperfec-
tions, or other practical problems, no transfer payments are allowed for, so that
the allocation cannot be relied upon the usual market mechanism.4 Nonethe-
less, it is desirable to allow students to choose their most preferred school if this
school is available. For example, consider the public school match currently in
use in Boston. Each school has a fixed maximum capacity of students it can
accommodate, which is determined by the Boston school department. Stu-
dents are given the opportunity to choose a school in kindergarten, and then
in first, sixth and ninth grade. So, clearly, the allocation or matching prob-
lem is dynamic. Moreover, it is reasonable to assume that students (and their
parents) are uncertain about the intensity of their future preferences for the
various schools when taking their decision in the presence. These features are
the main ingredients of the model we analyze. We capture the dynamic nature
by assuming that students go to school for two periods. There are two schools,
or two types of schools, each type with a fixed capacity, and individuals face
uncertainty about their future utility when applying to a school in the presence.
Related literature The present paper relates to the literatures on matching
theory and mechanism design without monetary transfers. A classic reference
for matching is Roth and Sotomayor (1990). Hatfield and Milgrom (2005) gen-
eralize some of the key results in matching markets. Static house allocation
problems with ordinal preferences have been studied by, among others, Ab-
dulkadiroglu and Sonmez (1998, 1999) and McLennan (2002). Roth, Sonmez,
and Unver (2004) use von-Neumann-Morgenstern utilities to assess the poten-
tial welfare gains from an improved matching mechanism for kidney transplants.
However, their mechanism does not elicit or use cardinal information. Simi-
larly, Roth, Sonmez, and Unver (2005b) analyze stochastic exchanges without
inducing revelation of cardinal utility information. Course bidding at business3See Abdulkadiroglu and Sonmez (2003), Abdulkadiroglu, Pathak, Roth, and Sonmez
(2005) and Abdulkadiroglu, Pathak, and Roth (2005). Though the model is meant to capturethe essential features of school assignments, it is also appropriate in other environments, mostimportantly perhaps, for the problem of queue management. Consider e.g. a large organi-zation, like a firm or a university, and assume that in every period various subunits of theorganization (called clients) require the services of a central unit (the server). Assume furtherthat the server’s capacity is limited, so that in every period some clients will be rationed. Thisproblem is almost identical to the school assignment problem we study. The main differenceis that the continuum assumption is less likely to be appropriate in this setting than in theschool assignment problem.
4Practical concerns arise, for example, when allocating resources within an organizationsuch as a firm or a university. Though in principle it is possible to create markets and havemarket mechanisms work inside the organization, there can be quite compelling reasons (suchas too much or too little consumption) why this need not be very desirable.
1 INTRODUCTION 4
schools in a static setting is studied by Sonmez and Unver (2005). The appli-
cation of the mechanism design approach to public school choice was proposed
by Abdulkadiroglu and Sonmez (2003),5 and some of the practical problems
and design issues involved are discussed by Abdulkadiroglu, Pathak, Roth, and
Sonmez (2005) and Abdulkadiroglu, Pathak, and Roth (2005).6
Casella (2003), Abdulkadiroglu (2004) and Jackson and Sonnenschein (2004)
have started to study mechanisms without transfers that improve allocations
in terms of expected utility, either by linking decisions (Casella, Jackson and
Sonnenschein) or by giving individuals ”quasi-money” (Abdulkadiroglu). An
important difference to Casella’s paper is that we look at a problem with a
continuum of agents. This allows us to abstract from strategic interactions
between agents, whereas she analyzes voting equilibria when there are two or
three agents who face a repeated decision problem. A main contrast between
our paper and those of Jackson and Sonnenschein and Abdulkadiroglu is the
striking simplicity of our mechanism, which nevertheless has very desirable
welfare properties. Hortala-Vallve (2004) and Borgers and Postl (2004) consider
the possibility of achieving welfare gains by studying more elaborate voting
schemes than ”one man one vote”. Hortala-Vallve proves the impossibility of
attaining first-best, while Borgers and Postl show that nonetheless substantial
welfare gains, as a matter of fact, almost first-best solutions, are possible.7
The remainder of the paper is structured as follows. Section 2 introduces
basic concepts and illustrates the main idea of the paper with a simple example.
Section 3 analyzes the general two period two school model, where all individ-
uals have identical ordinal preferences, utility for the bad school is normalized
to zero and utility draws for the good school are independent over time. In
Section 4, we analyze the same model with the modification that students have
heterogenous ordinal preferences, i.e. some students prefer one school and some
the other. Section 5 introduces uncertainty about ordinal preferences in addi-
tion to uncertainty about preference intensity. This setting is natural when a
school is not good or bad per se, but rather specialized, say, either in languages5See also Boston Globe (2003).6See also Roth, Sonmez, and Unver (2005a) and Roth (2002). The former address design
problems for kidney exchange, the latter provides a detailed discussion of other applicationsand issues in economic design.
7The basic problem Casella, Jackson and Sonnenschein, Hortala-Vallve, and Borgers andPostl address is a collective decision problem with few (typically two or three) individualswho may differ both with respect to their ordinal preferences as well as the intensities of thesepreferences. The goal of all four papers is to find mechanism that improve ex ante expectedutility.
2 EXAMPLE 5
or in math, and when students are uncertain in period one about the skills they
are going to develop during that period. Section 6 extends the model to allow
for correlation of utility over time. In Section 7, we study the model when
switching from one school to another involves a dead weight loss cost. Section
8 summarizes the main results we obtain for the various models, and Section 9
concludes. A generalization that drops the normalization of utility for the bad
school, which is maintained throughout the paper, and an example are deferred
to the Appendix.
2 Example
For the purpose of illustration, we first consider a simple example. There is a
continuum of students whose total mass is two. There are two schools, or two
types of schools, A and B, each with a capacity to accommodate students of
mass one. Students go to school for two periods, t = 1, 2. Within a period, a
school is an indivisible good, but students can switch school from one period to
another at zero costs.
Preferences over schools are as follows. Denote by xkt the instantaneous
utility of a student when attending school k = A, B in period t. We assume
that xBt = 0 for all t and all students. Instantaneous utility for A, xAt, is
drawn randomly from the distribution G on [0,M ] with M > 0 and G(0) = 0,
where these draws are i.i.d. over time and across students.8 Because utility of
school B is always zero, we write x for the utility of school A for simplicity.
Denote by Eu the expected value of x. The realizations of the instantaneous
utilities are known by individuals before applying to a school within a given
period. Utility is additive over time, and there is no discounting, so that in
period one the expected utility of an individual who is sure to attend school A
in both periods and whose first period draw is x is x + Eu; Figure 1 illustrates
the timing. Also, we let F (x) be the measure of students whose utility draw
for school A in period one is no larger than x. Because students have mass two
in total, we have F (x) ≡ 2G(x).8The assumption that G(.) is a continuous function is made merely for expositional conve-
nience. All the main results would go through if the random variable is of the discrete type,as long as its support includes sufficiently many points; in particular, all the main results gothrough if the support contains the points (if they exist) xj with j = 1, .., 4 defined in theproof of Proposition 1 below.
2 EXAMPLE 6
Figure 1: Timing.
2.1 Mechanisms
We investigate the potential of using dynamic mechanisms in a school assign-
ment (or house allocation) problem with two periods and two schools. We
restrict attention to mechanisms that are simple in that the number of possible
messages is equal to the number of schools. This type of mechanism is particu-
larly apt for practical use since asking (and answering) the question ”To which
school do you want to apply?” is a very natural thing in a school assignment
problem. Though more complicated mechanisms exist that may be used in prac-
tical applications as well, the analysis of such mechanisms is beyond the scope
of the present paper. There are two motivations for this. First, it simplifies the
analysis. Second, as will be shown below, there are reasonable conditions under
which even a very simple mechanism achieves the optimal incentive compatible
allocation over both periods.
In particular, we consider and compare the following two types of mecha-
nisms.
Static (or Myopic) Mechanism Let each individual report his or her pref-
erences in period one. If there is excess demand for a school, individuals are
allocated randomly. The same procedure is repeated in period two.
We contrast the static (or myopic) mechanism with the dynamic mechanism.
Dynamic Mechanism Those students who apply to the less preferred school
in period one are given priority for their preferred school in period two. (It
is assumed that everybody knows which school is preferred at the time the
mechanism is run.) Students who apply to the preferred school in period one
have priority for this school in period one. If there is excess demand in any
period for some school, students with the same priority are allocated randomly.
2 EXAMPLE 7
2.2 Equilibrium
We now derive equilibrium under both mechanisms.
Equilibrium under the Static Mechanism Let µ2 ∈ [0, 2] be the measure
of other students an individual believes will apply to A in t = 2 (not necessarily
the correct measure). Denote by Uk(x) the expected utility of an individual
whose utility draw for school A is x and who applies for school k. Under the
static mechanism, ”apply to A” is a strictly dominant strategy in t = 2 since
UA(x) = min{
1,1µ2
}x > 0 = UB(x)
for all x > 0. Thus, the static mechanism induces a unique equilibrium in t = 2.
In this equilibrium, all individuals apply to A.
Consequently, the expected utility of an individual in t = 1 whose first
period utility draw is x and who believes that µ1 ∈ [0, 2] others apply to A in
t = 1 when applying to A (taking equilibrium behavior in t = 2 into account)
is
UA(x) = min{
1,1µ1
}x +
12Eu >
12Eu = UB(x),
where Eu is the expected utility for school A and where the strict inequality
holds for all x > 0. Thus, ”apply to A” is a strictly dominant strategy in both
periods.
Equilibrium under the Dynamic Mechanism Consider the expected util-
ity of an individual who believes µ others apply to A in period one (and who
correctly anticipates that in t = 2 all individuals will apply to A) after his
period one utility x is realized, i.e. at the interim stage:
UA(x) = min{
1,1µ
}x + max
{0, 1− 1
µ
}Eu
UB(x) = max{
0,1− µ
2− µ
}x + min
{1,
12− µ
}Eu.
So, an individual with utility x applies to A if and only if
UA(x)− UB(x) =(
min{
1,1µ
}−max
{0,
1− µ
2− µ
})x
+(
max{
0, 1− 1µ
}−min
{1,
12− µ
})Eu ≥ 0.
3 THE MODEL 8
It is easy to see that regardless of whether µ ≥ 1 or µ < 1, the individual
will apply to A if and only if x ≥ Eu. Thus, every individual has a strictly
dominant strategy, which is either ”apply to A” or ”apply to B”. Which strat-
egy is dominant depends on the first period utility realization x. Because the
equilibrium is in strictly dominant strategies, it is necessarily unique.
2.3 Uniform Distribution
For the purpose of illustration, we consider now an example where G is uni-
form on [0,M ]. Thus, Eu = M2 and F (Eu) = 1. In t = 1, all individuals
with x ≤ Eu apply to B, and all others apply to A. Note that maximizing
social welfare requires that the fifty percent of students with the most intense
preference for A are assigned to A, which is exactly what is achieved under the
dynamic mechanism. Thus, from a social point of view, the first-best allocation
is achieved in t = 1.
Quantifying Welfare Gains First, since all students with x ≤ Eu apply to
B in t = 1, there is no rationing (or excess demand for school A) and no demand
shortage for school B in period one. Second, because of this and because all
high utility students (i.e. all students with x > Eu) go to school A in t = 1,
the period one allocation maximizes social welfare. Overall welfare under the
dynamic mechanism is given by
WDM = E[x | x > M/2] + Eu =54M.
Since M > 0, this is greater than expected welfare under the static mechanism,
which is WSM = M . The expected welfare gain generated by the dynamic
mechanism is thus M4 , or twenty-five percent.
3 The Model
We now consider a more general model, in which utility for the good school, A,
is drawn from the distribution function G with support [0,M ] and where school
A can have any capacity α ∈ (0, 2). Capacity of school B is 2− α.9 As before,
we denote by x the instantaneous utility for A (which is private information)
and by Eu the expected value of x and we let F (x) ≡ 2G(x) be the mass of
students with utility draw no larger than x. The median value of x is denoted9Aside from the fact that capacities of schools may, in general, vary, the exercise of allowing
for varying capacities is motivated by the model of Section 4, where students differ with respectto their preferred school.
3 THE MODEL 9
by m, i.e. G(m) = 12 . Notice that F (m) = 1. Utility for school B is zero
for all individuals and both periods, and there are no costs of switching from
one school to the other after period one. All of these assumptions are common
knowledge.
The assumptions of two schools is maintained throughout the paper. The
assumption of homogenous and time invariant ordinal preferences, i.e. that all
students have an ordinal preference for school A in both periods, is relaxed in
Section 4. Note that time invariant ordinal preferences imply certainty about
future ordinal preferences. This assumption will be relaxed in Section 5. The
assumption that switching from one school to another is costless is relaxed in
Section 7. The normalization of utility for school B is dropped in the Ap-
pendix, where we show that the normalization does not affect the results in any
qualitatively important way.
We begin with the definition of a cutoff equilibrium.
Definition 1 x∗ ∈ (0,M) is an (interior) equilibrium cutoff point if given the
behavior of all others it is optimal
- for every student with x < x∗ to apply to B,
- for every student with x ≥ x∗ to apply to A, and
- µ = 2− F (x∗).
We are now ready to state one of the main results of this section, which
is that under fairly general conditions a cutoff equilibrium under the dynamic
mechanism exists.
Proposition 1 For any α ∈(0, 2M
Eu+M
)and any G with full support on [0,M ]
and G(0) = 0, there is an interior cutoff equilibrium. A sufficient condition for
the existence of an interior cutoff equilibrium for any α ∈ (0, 2) is Eu = m.
Notice that 2MEu+M > 1 for any non-degenerate distribution G.
Proof : The proof consists of three steps. In step 1, we derive the four cases
that have to be distinguished and the necessary conditions for an equilibrium
for each case. In step 2, we construct an equilibrium for any α ≤ 1 and in step
3, we construct an equilibrium for all α > 1.
Step 1: Consider Figure 2 to see that as a function of α and µ, which
remains to be determined in steps 2 and 3, there are four cases that can occur.
Case 1 : Assume 2 − µ < α < µ. In this case, the number of applicants to
B in t = 1 is smaller than the capacity of A, which in turn is smaller than the
number of applicants to A. For this case to occur in a cutoff equilibrium, the
3 THE MODEL 10
Figure 2: Four cases.
following must hold for some x:
UA(x) =α
µx +
α− (2− µ)µ
Eu = Eu = UB(x)
⇔x∗ =
2− α
αEu.
Clearly, ∂UA∂x = α
µ > 0 = ∂UB∂x .
Case 2 : Assume µ < α < 2−µ. Then, the capacity of B is smaller than the
number of applicants to B in t = 1. Consequently, no applicant to A in t = 1
will get into A in t = 2.
UA(x) = x =α− µ
2− µx +
α
2− µEu = UB(x)
⇔x∗ =
α
2− αEu.
It is easy to see that ∂UA∂x = 1 > α−µ
2−µ = ∂UB∂x .
Case 3 : Assume α < min{µ, 2− µ}. Note that α < µ ⇔ 2− α > 2− µ, i.e.
there are less applicants to B in t = 1 than seats in B. Moreover, α < 2 − µ
implies that there are more applicants to B in t = 1 than the capacity of A can
3 THE MODEL 11
accommodate. Consequently, none of those who apply to A in t = 1 will get
into A in t = 2.
UA(x) =α
µx =
α
2− µEu = UB(x)
⇔x∗ =
µ
2− µEu.
Obviously, ∂UA∂x = α
µ > 0 = ∂UB∂x .
Case 4 : Assume α > max{µ, 2− µ}. The condition α > 2− µ implies that
the number of applicants to B in t = 1 is smaller than the capacity of A.
UA(x) = x +α− (2− µ)
µEu =
α− µ
2− µx + Eu = UB(x)
⇔x∗ =
2− µ
µEu.
It is easy to check that ∂UA∂x = 1 > α−µ
2−µ = ∂UB∂x .
These conditions and cases can be summarized as follows:
x1 =2− α
αEu 2− µ ≤ α ≤ µ (Case 1)
x2 =α
2− αEu µ ≤ α ≤ 2− µ (Case 2)
x3 =µ
2− µEu α ≤ min{µ, 2− µ} (Case 3)
x4 =2− µ
µEu α ≥ max{µ, 2− µ} (Case 4),
where we have dropped the ”star” and used subscripts to indicate the respective
cases. Second, replace µ by 2− F (xi) for i = 1, .., 4 to get:
x1 =2− α
αEu F (x1) ≤ α ≤ 2− F (x1)
x2 =α
2− αEu 2− F (x2) ≤ α ≤ F (x2)
x3 =2− F (x3)
F (x3)Eu α ≤ min{F (x3), 2− F (x3)}
x4 =F (x4)
2− F (x4)Eu α ≥ max{F (x4), 2− F (x4)}.
Note that ∂x1∂α < 0 and ∂x2
∂α > 0.
Step 2: Another necessary condition for some x∗ and a given µ to constitute
an equilibrium is that x∗ and µ must be consistent, i.e. it must be that µ =
2− F (x∗).
3 THE MODEL 12
So as to show that for α ∈(0, 2M
M+Eu
)there is always an equilibrium with an
interior cutoff point, it is useful to distinguish the cases with α ≤ 1 and α > 1.
We begin with the former and show that such a pair x∗ and µ = 2 − F (x∗)always exists, which then proves existence.
For α ≤ 1, the strategy of the proof is to start with an x close to the median
such that the restriction α ≤ min{F (x), 2−F (x)} of a case 3 equilibrium is met.
If this is not an equilibrium (i.e. if for no such x, x = x3 holds), then we can
either decrease x until we have an x that satisfies all of the restrictions of a case 1
equilibrium, or we can increase x until we have a case 2 equilibrium. For α > 1,
the strategy of the proof is completely analogous, except that we start with an
x close to the median that satisfies the restrictions α ≥ max{F (x), 2 − F (x)}of a case 4 equilibrium.
Recall that m denotes the utility of the median, i.e. F (m) = 1, and consider
case 3. Clearly, for α ≤ 1 there is always an x such that α ≤ min{F (x), 2 −F (x)}, since we can always choose x = m. If in addition, x = 2−F (x)
F (x) Eu holds,
we have an equilibrium. So, assume either x > 2−F (x)F (x) Eu or x < 2−F (x)
F (x) Eu for
all x for which the restriction α ≤ min{F (x3), 2 − F (x3)} holds. We consider
the first case first. Start with an x such that α ≤ min{F (x), 2 − F (x)} holds
and decrease x until x = x with x > 2−F (x)F (x) Eu and α = F (x) < 2− F (x). The
last inequality holds because initially 2−F (x) ≥ α holds and as x decreases to
x, 2 − F (x) > α follows. Note that x > 2−αα Eu. Now decrease x further until
x = 2−αα Eu. As x < x implies F (x) < α < 2−F (x), it follows that we have an
equilibrium of the form described in case 1.
Notice that x1 = 2−αα Eu > 0 for any α < 1. The only concern for the
existence of an interior equilibrium is thus that x1 > M . However, since x1 is
only needed to prove equilibrium existence when x3 is not an equilibrium cutoff
point and because in this case x1 is known to be smaller than some x < M , we
know that x1 < M holds.
So as to complete the case where α ≤ 1, assume now that for any x that
satisfies the restriction α ≤ min{F (x), 2 − F (x)}, x < 2−F (x)F (x) Eu holds. Start
with any such x and increase it until x equals x, which is such that x < 2−F (x)F (x) Eu
and α = 2 − F (x). Note that 2 − F (x) = α < F (x), since initially both
F (x) and 2 − F (x) were larger than α and since the term 2 − F (x) decreases
when x increases. Note also that x < α2−αEu. So, increase x further until
x = α2−αEu. This is clearly an equilibrium of the type analyzed in case 2, with
2− F (x) < α < F (x).
Notice that x2 = α2−αEu ≤ Eu for any α ≤ 1 and that x2 is positive. Thus,
3 THE MODEL 13
Figure 3: A fix point x3 always exists.
whenever x2 is an equilibrium cutoff point for α ≤ 1, it is an interior cutoff
point.
Step 3: The reasoning for the case with α > 1 is almost completely anal-
ogous. Consider an x such that α > max{F (x), 2− F (x)}. Because α > 1, we
know that such x’s always exist if we choose them close enough to m. If in ad-
dition for one such x, x = F (x)2−F (x)Eu holds, we have an equilibrium of the case 4
type. So, assume that no such x exists, i.e. whenever α > max{F (x), 2−F (x)}is satisfied, we either have x > F (x)
2−F (x)Eu or x < F (x)2−F (x)Eu. Consider first the
case where x > F (x)2−F (x)Eu and the restriction α ≥ max{F (x4), 2−F (x4)} is sat-
isfied. As we decrease x, F (x) decreases and 2−F (x) increases. Since initially
max{F (x), 2−F (x)} < α, the constraint that will become binding for some suf-
ficiently small x is α = 2−F (x) > F (x). Assume that x > F (x)2−F (x)Eu = 2−α
α Eu.
Clearly, as we decrease x further until x = 2−αα Eu, F (x) < α < 2−F (x) holds,
and we have an equilibrium of the case 1 type. Because we reach x1 by de-
creasing x, starting from some x < M , we know that x1 is an interior cutoff
point.
Finally, consider the case where for any x such that α > max{F (x), 2−F (x)}holds, we have x < F (x)
2−F (x)Eu. Increase x until α = F (x) > 2 − F (x). Note
that x < F (x)2−F (x)Eu = α
2−αEu. Increase x further up to x = α2−αEu. Since
F (x) > α > 2−F (x), we have an equilibrium of the type considered under case
2, provided x2 < M , which is implied by the assumption α < 2MM+Eu .
That Eu = m is a sufficient condition for an interior cutoff equilibrium to
exist for any α ∈ (0, 2) follows directly by plugging x3 = m and x4 = m into
the conditions x3 = 2−F (x3)F (x3) Eu and x4 = F (x4)
2−F (x4)Eu and the corresponding
restrictions α ≤ min{F (x3), 2−F (x3)} and α ≥ max{F (x4), 2−F (x4)}, which
yields x3 = x4 = Eu and min{F (x3), 2−F (x3)} = max{F (x4), 2−F (x4)} = 1.
3 THE MODEL 14
Figure 4: A fix point x4 may fail to exist.
Thus, for α ≤ 1, case 3 is an equilibrium, and otherwise case 4 is an equilibrium.
¥Notice that an interior fix point x3 = 2−F (x3)
F (x3) Eu always exists, as illustrated
in Figure 3. An interior fix point x4 = F (x4)2−F (x4)Eu, on the other hand, need not
exist because F (.) may be such that F (x)2−F (x)Eu > x for all x > 0; see Figure 4.
This happens, for example, with G(x) = x12 for 0 ≤ x ≤ 1, then Eu = 1
3 andF (x)
2−F (x)Eu = G(x)1−G(x)Eu > x for all x ∈ [0, 1].10
Not only the restriction Eu = m, but also the restriction α < 2MM+Eu in the
first part of the proposition is only sufficient. In Appendix B, we briefly illus-
trate this ’sufficiency without necessity’ with a simple example. We show also
that the restrictions have grip by providing an example where an equilibrium
does not exist when both restrictions are violated.
It is also worth mention that the possibility that a cutoff equilibrium does
not exist depends on the normalization of utility for school B in the follow-
ing sense. As shown in Appendix A, if utilities for school B and A, xB and
xA, are drawn independently from the distributions G[0,M ] and G[M, 2M ],
respectively, then a case 4 equilibrium exists for any α ≥ 1. Thus, without the
normalization, the restriction on G that Eu = m can be dispensed with.11
3.1 Welfare Properties
Proposition 1 asserts the existence of a cutoff equilibrium under fairly general
conditions. We are now going to discuss the welfare properties of these cutoff
equilibria.10One of the reasons why a fix point x4 does not exist in this case is that at x = 0, the
derivative of G(x)1−G(x)
Eu is infinite.11Of course, some symmetry is contained in the assumption that xB and xA are drawn from
the same distribution G with disjoint supports.
3 THE MODEL 15
Proposition 2 In any cutoff equilibrium, every individual expects a higher util-
ity at the interim stage in t = 1 under the dynamic mechanism than under the
static mechanism.
Proof : Denote by U(x) the expected interim utility of an individual under
a static mechanism, where everybody always applies to A. That is, U(x) ≡α2 x+ α
2 Eu. Given µ, the utility of applying to A under the dynamic mechanism
is
UA(x) = min{
1,α
µ
}x + max
{0,
α− (2− µ)µ
}Eu
and the utility of applying to B is
UB(x) = max{
0,α− µ
2− µ
}x + min
{1,
α
2− µ
}Eu.
Case 1 : If 2 − µ < α < µ, then UA(x) = αµx + α−(2−µ)
µ Eu ≥ α2 (x + Eu) =
U(x) ⇔ x ≥ 2−αα Eu. On the other hand, UB(x) = Eu > α
2 (x + Eu) = U(x) ⇔x < 2−α
α Eu.
Case 2 : If µ < α < 2 − µ, then UA(x) = x ≥ α2 (x + Eu) = U(x) ⇔ x ≥
α2−αEu. For UB(x) > U(x), we need α−µ
2−µ x+ α2−µEu > α
2 (x+Eu) ⇔ x < α2−αEu.
Case 3 : If α < min{µ, 2 − µ}, then UA(x) = αµx ≥ α
2 (x + Eu) = U(x) ⇔x ≥ µ
2−µEu. For UB(x) > U(x), we need α2−µEu > α
2 (x + Eu) ⇔ x < µ2−µEu.
Case 4 : If α > max{µ, 2−µ}, then UA(x) = x+ α−(2−µ)µ Eu ≥ α
2 (x+Eu) =
U(x) ⇔ x ≥ 2−µµ Eu. For UB(x) > U(x), we need UB(x) = α−µ
2−µ x + Eu >α2 (x + Eu) = U(x) ⇔ x < 2−µ
µ Eu. ¥For the static mechanism, it is immaterial whether it is a one period or
a two period problem. At the interim stage in t = 1, the expected utility of
an individual with utility draw x is U = α2 x + α
2 Eu. By Proposition 2, one
can ask individuals at the interim stage whether they prefer the dynamic or
the static mechanism. Under our assumptions, they will all prefer the dynamic
mechanism. Moreover, because interim expected utility under the dynamic
mechanism exceeds interim expected utility under a static mechanism for every
individual, we have also shown:
Corollary 1 Ex ante expected utility under the dynamic mechanism is larger
than under a static mechanism.
Overall Second-best Welfare Next we show that whenever the first-best
allocation is achieved in t = 1 under a dynamic mechanism, then this mecha-
nism achieves the second-best allocation over both periods, i.e. implements the
optimal incentive compatible allocation overall.
3 THE MODEL 16
Lemma 1 In the two-period game where all students prefer A to B, the optimal
incentive compatible second period allocation is a random allocation.
Proof In t = 2, the game reduces to a static game. Thus, all individuals have
a dominant strategy to apply to A (or to report the highest utility if asked to
report utilities). Consequently, based on period two reports, the allocation can
only be random. On the other hand, because of the i.i.d. assumption, period
one reports cannot be informative about period two utilities. Thus, there is no
way to improve upon a random allocation in period two. ¥For example, the dynamic mechanism establishes first-best in t = 1, and
second-best in t = 2 when both schools have equal capacities and when Eu = m.
Consequently, no other mechanism can do better. Thus:
Proposition 3 If the dynamic mechanism establishes first-best in period one,
it establishes second-best overall.
3.2 Unique Equilibrium vs. Multiple Equilibria
The following proposition characterizes completely the sets of equilibria under
the dynamic mechanism.
Proposition 4 Two cases occur:
• For α ≤ 1, there is a unique equilibrium under the dynamic mechanism.
This equilibrium is a cutoff equilibrium.
• For α > 1, all equilibria under the dynamic mechanism are either cutoff
equilibria, or all students apply to B.
Proof : First, we show that for any α ∈ (0, 2), there is no equilibrium where
all x ∈ [x, x] do the same (e.g. apply to A), while some x′ < x and some x′′ > x
do the converse (i.e. apply to B), where 0 < x ≤ x < M . Therefore, all
equilibria will be either cutoff or such that all individuals do the same. Second,
we show that for any α ∈ (0, 2) there is no equilibrium where all apply to A.
Moreover, for α ≤ 1, there is no equilibrium where all apply to B. For α > 1,
on the other hand, it cannot be ruled out in general that all apply to B. Third,
we deal with the question of multiple cutoff equilibria. This multiplicity may
occur for α > 1, but not for α ≤ 1.
Step 1: Denote by Pt(k) the probability of getting into school A in period
t when applying to school k in period one, k = A, B and t = 1, 2. Assume first
3 THE MODEL 17
that all x ∈ [x, x] apply to A. Then, we must have
UA(x) = P1(A)x + P2(A)Eu ≥ P1(B)x + P2(B)Eu = UB(x)
⇔[P1(A)− P1(B)]x ≥ [P2(B)− P2(A)]Eu.
Because it follows from our assumptions that at least F (x) > 0 individuals have
priority over applicants to A in period two, P2(B) > P2(A) follows. Therefore,
the right-hand side is positive, and consequently, the left-hand side must be
positive, too, implying P1(A) > P1(B). Therefore, ∂UA∂x > ∂UB
∂x > 0. Thus, the
stipulated behavior cannot be an equilibrium.
If all x ∈ [x, x] apply to B, the above inequality must be reversed. But
then, all x < x strictly prefer to apply to B, too, thus yielding the desired
contradiction.
Step 2: Assume all apply to A, i.e. µ∗ = 2. But then, for any α ∈ (0, 2),
UB(x) = Eu >α
2(x + Eu) = UA(x)
for x sufficiently close to zero.
If α ≤ 1, then there is no equilibrium where all apply to B because
UA(x) = x >α
2(x + Eu) = UB(x)
for x > α2−αEu, which for α ≤ 1 holds for any x > Eu, and thus in particular
for x = M . On the other hand, if α > 1, then α2−αEu > M may be the case.
Thus, for α > 1 all applying to B can be an equilibrium.
Step 3: From the previous steps, we know that the only equilibria are cutoff
for α ≤ 1. We are now going to show that there is a unique cutoff equilibrium
in this case by showing that the conditions of cases 1 trough 4 are incompatible.
Note first that for α = 1, all four cases are equivalent, provided their condi-
tions are consistent with equilibrium. That is, x1 = x2 = Eu. So if case 1 and
2 are both consistent with equilibrium, it must be that F (x1) = F (x2) = 1.
Similarly, if case 3 or case 4 is consistent with equilibrium, the constraints for
each of these two cases must hold with equality, implying F (x3) = F (x4) = 1.
Consider therefore the case with α < 1. Trivially, there is no case 4 equilib-
rium for α < 1, so cases 1, 2 and 3 remain to be checked.
Next note that x1 = 2−αα Eu > α
2−αEu = x2 implies F (x1) > F (x2), which
in turn implies 2− F (x1) < 2− F (x2). Because case 1 requires α < 2− F (x1)
while case 2 requires 2− F (x2) < α, it follows that the two cases are mutually
3 THE MODEL 18
exclusive. Thus, we are left to check consistency of cases 1 and 3 and cases 2
and 3.
So as to see that cases 1 and 3 are incompatible, note first that α < F (x3)
implies 2−α > 2−F (x3), which in turn implies x1 = 2−αα Eu > 2−F (x3)
F (x3) Eu = x3
and hence F (x1) > F (x3). But α ≤ F (x3) < F (x1) ≤ α is a contradiction.
Hence, the two cases are mutually exclusive. In order to see that cases 2 and
3 are incompatible, observe that α < 2 − F (x3) implies 2 − α > F (x3), which
in turn implies x3 = 2−F (x3)F (x3) Eu > α
2−αEu = x2. Hence, F (x3) > F (x2) ⇔2− F (x3) < 2− F (x2). But for case 3, α ≤ 2 − F (x3) must hold, whereas for
case 2, 2 − F (x2) ≤ α has to be satisfied, which with 2 − F (x3) < 2 − F (x2)
yields the desired contradiction.
Finally, turn to the case α > 1. Clearly, a case 3 equilibrium cannot occur
now. However, case 1 and case 2 are now not mutually exclusive since for
α > 1, x1 < x2 follows, implying F (x1) < F (x2) ⇔ 2− F (x1) > 2− F (x2). So,
depending on F (.) and α, F (x1) < α < 2− F (x1) and 2− F (x2) < α < F (x2)
can both hold. Moreover, either case can be consistent with case 4, depending
again on F (.) and α. ¥So as to develop some intuition and gain a better understanding of the wel-
fare implications of different cutoffs, an example with multiple cutoff equilibria
is instructive.
Welfare Across the Different Equilibria We now consider an example
where utility is drawn from the uniform distribution G(x) = xM on [0,M], so
that F (x) ≡ 2G(x) = 2xM and Eu = m = M
2 . For the four cases we have:
• Case 1 : x1 = 2−αα
M2 and F (x1) = 2−α
α .
• Case 2 : x2 = α2−α
M2 and F (x2) = α
2−α .
• Case 3 : x3 = M2 because Eu = m, so that F (x3) = 1.
• Case 4 : x4 = M2 because Eu = m, so that F (x4) = 2− F (x4) = 1.
For α < 1, case 3 is an equilibrium, and from Proposition 4 we know that it
is the unique equilibrium. For α > 1, case 4 is clearly an equilibrium. Case
1 is an equilibrium whenever F (x1) < α < 2 − F (x1) ⇔ 2−αα < α < 3α−2
α ,
which holds for all α ∈ (1, 2). A necessary condition for a case 2 equilibrium
is 2 − F (x2) < α < F (x2) ⇔ 4−3α2−α < α < α
2−α , which holds for all α ∈ (1, 2).
However, for case 2 to be an interior cutoff, it must also be the case that x2 < M ,
which requires α < 43 . Thus, for α ∈ (1, 4
3), there are three cutoff equilibria.
3 THE MODEL 19
Moreover, for α2−αEu > M ⇔ α > 4
3 , there is also an equilibrium where all
apply to B.
Clearly, under the equilibrium where all apply to B welfare is lowest because
this equilibrium induces just a random assignment in both periods. More inter-
esting is a comparison of welfare across the different cutoff equilibria. Welfare
is not monotonously increasing in the cutoff x∗. The reason is that though a
higher x∗ implies a greater efficiency among those who apply to A, a higher x∗
eventually also means more inefficiency among those who apply to B and who
eventually are assigned to A. Having fewer applicants for A (or equivalently,
having a higher cutoff) is only desirable if there is excess demand for A. If, in
equilibrium, the number of applicants is already reduced below A’s capacity,
then it seems desirable to have a lower cutoff, implying more applicants to A,
because the random assignment among those who apply to B is a source of
inefficiency as well.
Denote by g(x) the density of G(x). For the uniform distribution we have
g(x) = 1M . We focus on the case with α > 1, for otherwise, there is a unique
equilibrium. Since expected welfare in period two is the same for all allocations,
we only report period one welfare. We compute the ex ante expected utility of
an individual and denote by Wi the first-period welfare in case j with j = 1, 2, 4.
• W1 = α2−F (x1)
∫ Mx1
xg(x)dx = 2+α8 M
• W2 =∫ Mx2
xg(x)dx + α−(2−F (x2))F (x2)
∫ x2
0 xg(x)dx = 4−α8 M
• W4 =∫ M
M2
xg(x)dx + (α− 1)∫ M
20 xg(x)dx = 2+α
8 M .
Note that W1(α) = W4(α). Moreover, for α = 1, welfare is the same across all
equilibria, but it increase in α for case 1 and 4 while it decreases in α in case
2. Thus, for α ∈ (1, 43),
W1 = W4 > W2.
3.3 Many Periods
We conclude this section with a very brief discussion of what happens if there
are more than two periods. Denote by WDM(2) the overall welfare achieved in
the equilibrium under the dynamic mechanism in the two period model, under
the assumption that such an equilibrium exists. Without much loss, assume
that the number of periods T is even12 and denote by WDM(T ) equilibrium
welfare under a dynamic mechanism for the T period model.12If not, the following statements are true for the model confined to the first T ≡ T − 1
periods.
4 HETEROGENEITY IN ORDINAL PREFERENCES 20
Proposition 5 Under the above assumptions, there is a dynamic mechanism
that induces a cutoff equilibrium such that
WDM(T ) ≥ T
2WDM(2)
holds.
Proof : The existence of a dynamic mechanism that induces a cutoff equilibrium
follows immediately from the existence of such an equilibrium under a dynamic
mechanism for the two period model: Repeat the two period mechanism T/2
times. The resulting equilibrium welfare will be WDM(T ) = T2 WDM(2). How-
ever, in general the longer time horizon may allow for even better mechanisms.
Therefore, WDM(T ) ≥ T2 WDM(2) holds. ¥
4 Heterogeneity in Ordinal Preferences
We now consider four extensions of the basic model. First, we study the model
under the assumption that individuals are heterogenous with respect to ordinal
preferences. Second, we analyze a simple model when students are uncertain
about which school they will prefer in the future. Third, we look at the case
when utilities are correlated across time. Lastly, we analyze what happens when
there is some inertia in the sense that students dislike switching from one school
to another.
4.1 Mechanisms when Ordinal Preferences are not Known
So far, we have assumed that all students prefer school A to B, though the
intensity with which they do so may differ across time. We now replace this
assumption by the alternative that some students have an ordinal time-invariant
preference for school A and others an ordinal and time-invariant preference for
school B. These preferences are private information, so that even the fraction
of students preferring A to B is not known. These assumptions are natural
when a school is not good or bad per se, but rather specialized, say, either in
languages or in math, and when it is not known how many students prefer to
specialize in math or languages, respectively.
There are two schools A and B with capacities α and 2 − α, respectively,
and two periods. Let xkt be distributed according to G(.) with support [0,M ]
and let x−kt = 0 for all students, where −k means ”not k”, k = A,B and
t = 1, 2. We first describe an augmented dynamic mechanism that allows to
infer the true preferences before individuals are asked to apply to the good or
5 UNCERTAINTY ABOUT ORDINAL PREFERENCES 21
bad school in period one.13 Second, we derive equilibrium behavior under this
mechanism.
Dynamic Two Phase Mechanism
Phase 1: Ask all individuals i to report their ordinal preferences Âi. If i’s
preferred school is available (i.e. if there is no excess demand for this school), i
is assigned to this school for both periods. Individuals who prefer a school with
excess demand enter phase 2.
Denote the true number of students preferring A to B by a. Without loss,
assume that A is the school with excess demand, i.e. a > α.
Phase 2: The mechanism announces the number of individuals who re-
ported that they preferred the school with excess demand. Then the dynamic
mechanism of Section 3 is applied, i.e. if a student now applies to the less
preferred school, he’ll have priority for the preferred school in t = 2.
4.2 Equilibrium under the Two Phase Mechanism
Lemma 2 Truth telling in phase 1 is a strictly dominant strategy.
Proof : Consider an individual i with A Âi B and let a be the measure of
other individuals who report that they prefer A to B. If a < α, then by telling
the truth he gets his first-best (i.e. into A in both periods, which is obviously
better than lying). If a ≥ α, there is too little demand for B. Thus, by lying he
gets into B in both periods, which is the worst that can happen to him. Thus,
truth telling is strictly dominant. ¥The lemma implies that a = a in equilibrium. Consequently, free capacity
of school B is a − α > 0. Without loss, we can normalize a = 2 and α = 2aα.
Therefore, we have shown:
Proposition 6 The game in phase 2 is equivalent to the one studied in Section
3, where capacity of school A is α.
5 Uncertainty about Ordinal Preferences
This section contains a generalization to the case where there is uncertainty both
about ordinal and cardinal future preferences (as opposed to only the cardinal13Put differently, under this mechanism ordinal and cardinal preference statements are made
in separate steps. The motivation is the same as the one of Sonmez and Unver (2005), whoobserve that order and intensity of preferences cannot be both inferred when observing onlya single variable (”bids”).
5 UNCERTAINTY ABOUT ORDINAL PREFERENCES 22
component). This model is appropriate when each school is specialized in, say,
languages or math, and when students are uncertain in period one about the
skills they are going to develop. It is shown for the example of a uniform
distribution that previous results basically carry over.
5.1 Assumptions
There are two periods, and two schools A and B with equal capacities (i.e.
α = 1). Total mass of students is two, and instantaneous utilities for school
A and B are drawn independently from the distributions GA(x) and GB(x)
with support [M,M]. These draws are i.i.d. across students and time, and as
before, there is no discounting. We assume that school A is the good school in
the sense that EUA > EUB.
There are now students who prefer school B to A in one period and A to
B in the other period. Denote by γ the probability that a student prefers B to
A. That is,
γ ≡ Pr(xB > xA) =∫ M
M
∫ y
MdGA(x)dGB(y) =
∫ M
MGA(y)dGB(y)
and Pr(xB ≤ xA) = (1 − γ). Consequently, some students who applied to
B in t = 1 will prefer school B to A in period two. Therefore, the dynamic
mechanism is as follows:
Dynamic Mechanism: If you apply to B in t = 1, you are in B, provided
there is enough capacity. In t = 2, everybody can again apply to A or B.
Those who applied to B in t = 1 have priority over all those who applied to A,
whatever school they apply to in t = 2.
Before we can proceed with the equilibrium analysis, some further definitions
are needed. The expected utility of A of a student who wants to go to A, i.e.
conditional on xA ≥ xB, is
EUA ≡ E[xA | xA ≥ xB] =∫ M
M
∫ xA
MxA
dGB(xB)dGA(xA)1− γ
,
where 11−γ is the conditioning factor. Next, denote by EUB the expected utility
of school B, conditional on xA > xB, which is given by
EUB ≡ E[xB | xA ≥ xB] =∫ M
M
∫ M
xB
xBdGA(xA)dGB(xB)
1− γ.
This will be of importance because an individual who would like to go to school
A may be assigned to B despite xA > xB, in which case his expected utility is
EUB.
5 UNCERTAINTY ABOUT ORDINAL PREFERENCES 23
5.2 Equilibrium for the Uniform Distribution
Assume GA and GB are uniform on [0, 2M ] and [0,M ], so that their densities
are gA = 12M for 0 ≤ x ≤ 2M and gB = 1
2M for 0 ≤ x ≤ M , respectively. The
unconditional expected utilities then are M and M2 . The probability of having
a preference for school B is then easily seen to be
γ =∫ M
0
∫ y
0
12M2
dxdy =14.
The joint density of xA and xB is 12M2 and the probability of xB ≥ xA is γ = 1
4 .
Hence, the joint conditional density of xA and xB, conditional on xA ≥ xB, is
g(xA, xB | xA ≥ xB) = 23M2 . Therefore, the relevant conditional expectations
are
EUA = E[xA | xA ≥ xB] =∫ M
0
∫ 2M
xB
xA
2M2
11− γ
dxAdxB =119
M
and
EUB = E[xB | xA ≥ xB] =∫ M
0
∫ 2M
xB
xB
2M2
11− γ
dxAdxB =49M.
Proposition 7 The game has an equilibrium in which individuals apply to A
in t = 1 if and only if x ≡ (xA, xB) is such that
xA ≥ xB + (1− 2γ)[EUA − EUB] = xB +718
M.
In this equilibrium, µ∗ = 109 .
Proof : Notice first that [EUA − EUB] = 79M and recall γ = 1
4 . Since all
individuals in the area above xA ≥ 718 + xB apply to A, it is easily seen that
their measure is 109 . So, assume this is an equilibrium. Then, the following
holds: µ∗ > 1 > (2 − µ∗) and 1 > (2 − µ∗)(1 − γ), so all who applied to B
in t = 1 will get into A if they want in t = 2. On the other hand, for γ < 12
the residual capacity of A in t = 2, 1 − (2 − µ∗)(1 − γ), does not permit to
accommodate all those who applied to A in t = 1 and who want to get into A
in t = 2 since γ < 12 ⇔ 1− (2− µ∗)(1− γ) < µ∗(1− γ). The cutoff equilibrium
condition thus reads
UA(x) =1µ∗
xA +(
1− 1µ∗
)xB + (1− γ)
×[1− (2− µ∗)(1− γ)
µ∗(1− γ)EUA +
(1− 1− (2− µ∗)(1− γ)
µ∗(1− γ)
)EUB
]+ γEUB
≥ xB + (1− γ)EUA + γEUB = UB(x).
Re-arranging and simplifying yields the condition in the proposition. ¥
5 UNCERTAINTY ABOUT ORDINAL PREFERENCES 24
5.3 Welfare Comparisons
Static mechanism Again, we restrict attention to period one welfare. Under
a static mechanism, all students with a preference for school A apply to this
school in t = 1. Since total mass of these students is 32 and since α = 1, one
third of them has to be rationed and sent to school B. On the other hand, since
γ < 12 , all students with a preference for B get into B for sure. Hence, so as to
derive welfare under a static mechanism, we first need to compute additionally
the expected utility of getting into B conditional on xB ≥ xA, denoted as
E[xB | xB ≥ xA]. The joint conditional density of xA and xB, conditional on
xB ≥ xA, is g(xA, xB | xB ≥ xA) = 2M2 . Consequently,
E[xB | xB ≥ xA] =∫ M
0
∫ xB
0xB
2M2
dxAdxB =23M.
Thus, expected welfare under a static mechanism, WSM , is
WSM = (1− γ)[23E[xA | xA ≥ xB] +
13E[xB | xA ≥ xB]
]+ γE[xB | xB ≥ xA]
=1318
M = 0.72M.
Next, we consider the equilibrium welfare under the dynamic mechanism.
Dynamic Mechanism Under the dynamic mechanism, only those indi-
viduals with xA ≥ xB + (1 − 2γ)[EUA − EUB
]= xB + 7
18M apply to school
A. The conditional expectations relevant for welfare are therefore E[xB | xA ≤xB + 7
18M ], E[xA | xA ≥ xB + 718M ] and E[xB | xA ≥ xB + 7
18M ]. The proba-
bility of xA ≥ xB + 718M is 5
9 . Thus, g(xA, xB | xA ≥ xB + 7
18M)
= 910M2 , and
g(xA, xB | xA ≤ xB + 7
18M)
= 98M2 . Hence
E[xB | xA ≤ xB +718
M ] =∫ M
0
∫ xB
0xB
98M2
dxAdxB =1932
M
E[xA | xA ≥ xB +718
M ] =∫ M
0
∫ 2M
xB
xA9
10M2dxAdxB =
1013720
M
E[xB | xA ≥ xB +718
M ] =∫ M
0
∫ 2M
xB
xB9
10M2dxAdxB =
1740
M.
Since the measure of students applying to A is 109 , which exceeds school A’s
capacity, one tenth of them will be rationed and sent to school B. Thus,
expected welfare under the dynamic mechanism, WDM , is given as
WDM =59
[910
E[xA | xA ≥ xB +718
M ] +110
E[xB | xA ≥ xB +718
M ]]
+49E[xB | xB ≥ xB +
718
M ] = 0.99M.
6 CORRELATIONS 25
Compared to the welfare under a static mechanism, this seems like a substantial
improvement. Now let us compare it with the first-best welfare.
First-best Under first-best, students with xA ≥ xB + M2 should be assigned
to school A, and all others should go to B. Note that because Pr(xA ≥ xB +M2 ) = 1
2 there is no rationing at either school, which obviously should be the
case under first-best. The probability of xA ≥ xB + M2 being 1
2 , the joint
conditional densities are 1M2 . The expected utility of getting into A, conditional
on xA ≥ xB + M2 is,
E[xA | xA ≥ xB +M
2] =
∫ M
0
∫ 2M
xB+M2
xA1
M2dxAdxB =
3524
M,
and the expected utility of getting into B, conditional on xA ≤ xB + M2 , is
E[xB | xA ≤ xB +M
2] =
∫ M
0
∫ xB+M2
0xB
1M2
dxAdxB =712
M.
Thus, first-best welfare is given as
WFB =12
[3524
M +712
M
]=
4948
M = 1.02M.
Comparing this number to the welfare achieved under the dynamic mechanism,
we see that welfare in period one under the dynamic mechanism is only three
percent less than first-best. So, adding the complication of uncertainty about
future ordinal preferences does not appear to have a big impact on the efficiency
properties of the dynamic mechanism.
6 Correlations
We now relax the assumption that utility draws are i.i.d. over time. We show
that even with positive correlation, expected welfare under the dynamic mecha-
nisms exceeds welfare under the static mechanism for any correlation coefficient
less than one. For perfectly positively correlated utility, welfare is the same un-
der the dynamic and under the static mechanism. Finally, we briefly analyze
the model with negative correlation.
6.1 Assumptions
There are two schools A and B, and total mass of students is two. As above,
capacity of A is α and capacity of B is 2−α. There are two periods, and utility
for school B is zero for both periods. Utility for A is greater than zero for all
individuals and both periods. (We will be more precise about how utility for
6 CORRELATIONS 26
A is generated shortly.) Denote by x the first period utility draw and by y the
utility for the second period.
6.2 Positive Correlation
So as to model positive correlation, we follow Jackson and Sonnenschein (2004,
Appendix 3, p. 45-6) and make the following assumptions. After period one
utility x is realized, period two utility y is
y = x
with probability ρ ∈ [0, 1] and drawn from the distribution G[0,M ] with prob-
ability (1− ρ). Thus,
E[y | x] = ρx + (1− ρ)Eu.
Existence for the Case with Positive Correlation We now show that a
result that is closely related to Proposition 1 holds for the case with positive
correlation.
Proposition 8 Provided α and ρ are such that ρ ≤ min{
α2−α ,
2−αα
M−Eu
M−Eu
},
there is an interior cutoff equilibrium under the dynamic mechanism. A suf-
ficient condition for the existence of an interior cutoff equilibrium for any
α ∈ (0, 2) and ρ ∈ [0, 1] is Eu = m.
Proposition 8 generalizes Proposition 1 to the case with positive correlation.
The only additional restriction we need to take care of is x1(ρ) > 0, which holds
if ρ < α2−α . On the other hand, x2(ρ) < M is guaranteed by ρ <
2−αα
M−Eu
M−Eu . This
is a generalization of the condition α < 2MEu+M of Proposition 1.14 Therefore,
the proof is analogous to the one of Proposition 1, which is why we only sketch
it.
Sketch of Proof : It is straightforward to derive the candidate equilibrium
cutoffs xi(ρ) with i = 1, .., 4:
x1(ρ) =2− α
α− ρ(2− α)(1− ρ)Eu F (x1) ≤ α ≤ 2− F (x1)
x2(ρ) =α
2− α− ρα(1− ρ)Eu 2− F (x2) ≤ α ≤ F (x2)
x3(ρ) =2− F (x3)
F (x3)− ρ(2− F (x3))(1− ρ)Eu α ≤ min{F (x3), 2− F (x3)}
x4(ρ) =F (x4)
2− F (x4)− ρF (x4)(1− ρ)Eu α ≥ max{F (x4), 2− F (x4)},
14To see this, set ρ = 0 in the condition of Proposition 8 to get the condition of Proposition1.
6 CORRELATIONS 27
where we have dropped ρ’s inside xi(.) on the right-hand side for notational
ease.
Next proceed as in the case with ρ = 0. That is, separate the problem into
the case with α ≤ 1 and α > 1 and consider the former first.
1. Choose x close to the median m, such that α ≤ min{F (x), 2 − F (x)}holds. If in addition x = 2−F (x)
F (x)−ρ(2−F (x))(1− ρ)Eu, we are done.
2. So assume it does not hold for any x that satisfies α ≤ min{F (x), 2 −F (x)}. Then, for all these x’s either (a) x > 2−F (x)
F (x)−ρ(2−F (x))(1 − ρ)Eu or (b)
x < 2−F (x)F (x)−ρ(2−F (x))(1− ρ)Eu holds.
Consider first (a). Decrease x until x, where F (x) = α and note that
x >2− α
α− ρ(2− α)(1− ρ)Eu =
2− F (x)F (x)− ρ(2− F (x))
(1− ρ)Eu.
So, decrease x further until x = 2−αα−ρ(2−α)(1 − ρ)Eu > 0 holds. That such a
2−αα−ρ(2−α)(1− ρ)Eu > 0 exists is guaranteed by the assumption ρ < α
2−α .
Consider now (b). Increase x until x = x, where α = 2 − F (x). Note that
x < α2−α−ρα(1− ρ)Eu. So, increase x further until x = α
2−α−ρα(1− ρ)Eu < M .
The inequality is satisfied because of the assumption ρ <2−α
αM−Eu
M−Eu , which is
satisfied for any α ≤ 1 since for these α’s2−α
αM−Eu
M−Eu > 1 holds.
3. Assume α > 1. Choose x close to m. Assume that for no such x,
x = F (x)2−F (x)−ρF (x)(1 − ρ)Eu holds for otherwise we are done. So, either (c)
x > F (x)2−F (x)−ρF (x)(1 − ρ)Eu or (d) x < F (x)
2−F (x)−ρF (x)(1 − ρ)Eu holds. Consider
first (c). Decrease x until x, where α = 2 − F (x) > F (x). Decrease x further
until x = 2−αα−ρ(2−α)(1 − ρ)Eu holds. Because for the larger x, we had α =
2 − F (x) > F (x), the restriction F (x) < α < 2 − F (x) is clearly satisfied.
Because of the assumption ρ < α2−α , x > 0 exists.
Finally, consider (d). Increase x in case the restriction is never satisfied
until x = x, where α = F (x). Increase x further until x = α2−α−ρα(1− ρ)Eu.
Like the conditions in Proposition 1, the condition ρ ≤ min{
α2−α ,
2−αα
M−Eu
M−Eu
}
in Proposition 8 is only sufficient. To see this, consider a distribution G satis-
fying Eu = m. Then, x3(ρ) = x4(ρ) = Eu for any ρ ∈ [0, 1]. Clearly, case 3 is
an equilibrium for any α ≤ 1, and case 4 is an equilibrium for any α > 1. Thus,
there will be an interior cutoff equilibrium for any α and ρ, including those for
which the condition is not met. ¥
Welfare with Positive Correlation Next we compare expected utility in
the interim stage under the dynamic mechanism with expected interim utility
6 CORRELATIONS 28
under a static mechanism. The following proposition is the analogue to, and a
generalization of, Proposition 2 for the case with positively correlated utilities.
Proposition 9 In the interim stage of a cutoff equilibrium with positive corre-
lation, expected utility under the dynamic mechanism is larger than under the
static mechanism.
Proof : The proof mimics the one for Proposition 2. As in Proposition 2, denote
by U(x) = α2 (x + E[y | x]) = α
2 ((1 + ρ)x + (1 − ρ)Eu) the expected utility in
the interim stage under the static mechanism. Under the dynamic mechanism,
we have:
Case 1 : UA(x) = αµx + α−(2−µ)
µ E[y | x] ≥ α2 (x + E[y | x]) = U(x) ⇔ x ≥
2−αα−ρ(2−α)(1−ρ)Eu = x1(ρ). On the other hand, UB(x) = E[y | x] ≥ α
2 (x+E[y |x]) = U(x) ⇔ x ≤ 2−α
α−ρ(2−α)(1− ρ)Eu = x1(ρ).
Case 2 : UA(x) = x ≥ α2 (x + E[y | x]) = U(x) ⇔ x ≥ α
2−α−ρα(1 − ρ)Eu =
x2(ρ). On the other hand, UB(x) = α−µ2−µ x + α
2−µE[y | x] ≥ α2 (x + E[y | x]) =
U(x) ⇔ x ≤ α2−α−ρα(1− ρ)Eu = x2(ρ).
Case 3 : UA(x) = αµx ≥ α
2 (x + E[y | x]) = U(x) ⇔ x ≥ µ2−µ−ρµ(1− ρ)Eu =
x3(ρ). On the other hand, UB(x) = α2−µE[y | x] ≥ α
2 (x + E[y | x]) = U(x) ⇔x ≤ µ
2−µ−ρµ(1− ρ)Eu = x3(ρ).
Case 4 : UA(x) = x + α−(2−µ)µ E[y | x] ≥ α
2 (x + E[y | x]) = U(x) ⇔ x ≥µ
2−µ−ρµ(1 − ρ)Eu = x4(ρ). On the other hand, UB(x) = α−µ2−µ x + E[y | x] ≥
α2 (x + E[y | x]) = U(x) ⇔ x ≤ µ
2−µ−ρµ(1− ρ)Eu = x4(ρ). ¥An immediate corollary to Proposition 9 is:
Corollary 2 Whenever an interior cutoff equilibrium exists with positive cor-
relation, ex ante expected utility under the dynamic mechanism exceeds ex ante
expected utility under a static mechanism.
6.3 Negative Correlation
Though the case with positive correlation may seem more relevant for real
world applications, there are also situations where period one and two utilities
may be negatively correlated. Consider the server-client example discussed in
the Introduction. If the server supports clients who experience stress due to
exogenous shocks (e.g. deadlines) and if these shocks are negatively correlated
over time (which is likely to be the case for deadlines), then there will be
negative correlation.
We capture negative correlation by assuming that y = M − x with prob-
ability ρ ∈ [0, 1] and drawn from G[0,M ] with probability (1 − ρ). Moreover,
6 CORRELATIONS 29
the distribution G[0,M ] from which utility for A is drawn is now assumed to
be symmetric and continuous, i.e. g(x) = g(M − x). This assumption makes
sure that the unconditional expectation of period two utility is the same as the
expectation of period one utility.
Existence Before we prove existence, some preliminary observations are help-
ful. We write EX [.] for the expectation taken with respect to the distribution
of x.
Claim: EX [x] = EX [M − x] = M2 when g(.) is symmetric.
Proof: First, EX [M − x] = M − EX [x], which is immediate. Second,
EX [x] = EX [M −x]. (Together with the first observation this then proves that
EX [x] = M2 .) To see that EX [x] = EX [M − x] is true, note that
EX [x] =∫ M
0xg(x)dx =
∫ M
0(M − x)g(M − x)dx =
∫ M
0(M − x)g(x)dx,
where the last equality holds because of symmetry. But∫ M0 (M − x)g(x)dx =
EX [M − x]. ¥Due to symmetry, it is also true that the median m is equal to M
2 = E[x] ≡Eu, where we drop the subscript X if there is no danger of confusion. Moreover,
E[y | x] = ρ(M − x) + (1− ρ)Eu = (1 + ρ)Eu− ρx, where the second equality
follows because M = 2Eu.
We can now state the general existence result, which is straightforward to
prove because G is symmetric.
Proposition 10 With negative correlation (and G symmetric), an interior
cutoff equilibrium always exists.
Proof : Consider the four cases of Proposition 1. Replace Eu by E[y | x] =
(1 + ρ)Eu− ρx and solve for the respective cutoff xi(ρ), i = 1, .., 4 to get
x1(ρ) =2− α
α + ρ(2− α)(1 + ρ)Eu F (x1) ≤ α ≤ 2− F (x1)
x2(ρ) =α
2− α + ρα(1 + ρ)Eu 2− F (x2) ≤ α ≤ F (x2)
x3(ρ) =2− F (x3)
F (x3) + ρ(2− F (x3))(1 + ρ)Eu α ≤ min{F (x3), 2− F (x3)}
x4(ρ) =F (x4)
2− F (x4) + ρF (x4)(1 + ρ)Eu α ≥ max{F (x4), 2− F (x4)}.
Because G is symmetric, Eu = m holds. Thus, for α ≤ 1 case 3 is always an
equilibrium, and for α > 1, case 4 is always an equilibrium. ¥
7 SWITCHING COSTS 30
Welfare Next we address interim expected utility.
Proposition 11 With negative correlation (and G symmetric), interim ex-
pected utility under the dynamic mechanism exceeds interim expected utility
under a static mechanism.
The proof is omitted because it is a one-to-one replication of the proofs of
Propositions 2 and 9. The only additional thing that one needs to keep in mind
is that M = 2Eu because the distribution G is symmetric. The proposition also
implies that ex ante expected welfare under the dynamic mechanism exceeds
expected welfare under a static mechanism.
Welfare with Equal Capacities Assume α = 1. Then, the cutoff in
period one is x1 = m, and we have first-best in period one:
WDM1 = E[x | x ≥ m] = WFB
1 .
In period two, welfare is
WDM2 = 2
∫ m
0(ρ(M − x) + (1− ρ)E[x]) g(x)dx,
where 2g(x) is the density conditional on x ≤ m. Thus,
WDM2 = (1− ρ)E[x] + 2ρ
∫ m
0(M − x)g(x)dx = (1− ρ)E[x] + ρE[x | x ≥ m].
Therefore, overall, welfare under the dynamic mechanism is
WDM (ρ) = (1 + ρ)E[x | x ≥ m] + (1− ρ)E[x],
which for ρ = 1 is equal to first-best welfare WFB = 2E[x | x ≥ m].
7 Switching Costs
In many settings, it is natural to assume that there is some kind of inertia. For
example, given that a student has already been assigned to a school, he or she
may be less inclined to go to the other school even if the other school is, a priori,
perceived as better. Similarly, when assigning tasks within an organization
or assigning houses to individuals, there may be economies of scale because
handing over the task or apartment from one individual to another is costly. So
as to take this possibility into account, we now introduce a switching cost s ≥ 0
that is borne by every student upon switching from one school to another.
However, before turning to the details of the model, a few comments are in
order. Clearly, any cost of switching, say, schools must be put into proportion
7 SWITCHING COSTS 31
with the utility generated by a school. Though this cost may be relatively high
when schools were to be switched, say, every year, it is arguably much smaller
when students switch from one school to another every five or six years, as is the
case under the Boston school assignment scheme discussed in the Introduction.
In particular, when periods are sufficiently long, switching costs may be close
to zero or even become negative.15
7.1 The Model with Switching Costs
The assumptions are now slightly modified to incorporate switchings costs. As
before, there are two periods, two schools A and B and students have homoge-
nous and time-invariant ordinal preferences for the two schools. Utility for
school B is normalized to zero, while utility for A is drawn from the distribu-
tion G with support [0,M]. Upon changing from one school to another, a student
bears a switching cost s ≥ 0.
As in the case with correlation, when assessing welfare one can no longer
restrict attention to the period one allocation. Students who are re-assigned in
t = 2 bear switching costs, which are a dead weight loss and hence work against
the desirability of our mechanism. On the other hand, in t = 2 some students
who are in B and who have low preference intensity for A prefer to stay in B,
which all else equal enhances the efficiency of the mechanism.
Introducing inertia in the form of switching costs does not only have the
effect of reducing the desirability of the dynamic mechanism. It also renders the
model much less tractable. We therefore proceed as follows. Instead of treating
the model in complete generality, we derive first the cutoff x∗ for an equilibrium
related to a case 2 equilibrium (as in Section 3) under the assumption that such
an equilibrium exists. Second, when G is uniform we derive sufficient conditions
for such an equilibrium to exist. Third, we discuss the welfare properties of the
dynamic mechanism and compare them with those for a static mechanism and
with first-best. Fourth, we show that for equal capacities and the distributions
G with Eu = m there is no case 3 and case 4 equilibrium.16 Finally, we briefly
discuss a modified dynamic mechanism that induces less switching and yields
larger welfare than the standard dynamic mechanism used so far.15This would be the case if parents feel that it is a good thing if their children learn to
adopt to a new environment every once in a sufficiently long while.16Equilibria of type 1 with switching costs, on the other hand, turn out to be very hard to
derive, which is why we chose not address them here.
7 SWITCHING COSTS 32
7.2 Case 2 Equilibrium Conditions
Recall that a case 2 equilibrium is such that µ < α < 2−µ ⇔ 2−µ > 2−α. In
this equilibrium, all students who apply to A in t = 1 will get into A in t = 1
and into B in t = 2 with certainty. With a positive cost of switching schools,
though, only the fraction 1−G(s) of the (2−µ) students in B want to get into
A because the students who have been assigned to B in t = 1 and whose utility
draw x is less than s in t = 2 prefer to stay in B. So as to preserve the above
mentioned properties of a case 2 equilibrium, the condition
(2− µ)(1−G(s)) > α (1)
has therefore to hold additionally, which will hold only if s is not too large. The
adjusted conditions for a case 2 equilibrium are thus
µ < α < 2− µ and (2− µ)(1−G(s)) > α.
Some additional notation is useful. Temporarily, let β denote the probability
of getting into A when applying to B and denote consequently by (1− β) the
probability of getting into B when applying to B in t = 1. Similarly, let δ and
(1− δ) be the probability of getting into A and B in t = 2, respectively, when
applying to B in t = 1.17 Also, denote by E[x | x > s] the expected utility for
school A conditional on the utility for A being greater than the switching cost
s, while the unconditional expected utility for school A is still denoted as Eu.
With this notation at hand, the expected utility of applying to B given
utility draw x in t = 1 is
UB(x) = β [x + δEu− (1− δ)s] + (1− β)δ(1−G(s)) (Eu[x | x > s]− s) , (2)
while the expected utility when applying to A is simply
UA(x) = x− s. (3)
Because in an (adjusted) case 2 equilibrium, those who apply to A get into
A for sure in t = 1 and into B for sure in t = 2, UA(x) = x − s follows
immediately. Equality (2) is also easily understood. With probability β, the
individual is assigned to A in t = 1, in which case he gets utility x in t = 1.
With probability δ, he will remain in A in t = 2, whence he derives an expected
utility of Eu. With probability (1 − δ), though, he has to switch schools in17Of course, these probabilities are endogenous and depend on the behavior of the individ-
uals under the mechanism. They will be expressed in terms of the primitives of the modelshortly.
7 SWITCHING COSTS 33
t = 2, which costs him s. This explains the first expression on the right-hand
side of (2). As to the second term, note first that the individual is assigned to
school B for both periods with probability (1−β)(1− δ), in which case he gets
zero utility but bears also no switching costs, which is why there are only two
terms on the right-hand side. With probability (1−β)δ the individual who was
in B in t = 1 switches to A in t = 2. His expected gross utility of being in
school A is Eu[x | x > s] since he only applies to school B in t = 2 when x > s.
In this case, though, he bears also the switching cost s, so that his net utility is
Eu[x | x > s] − s. Because this happens only with probability (1−G(s)), the
expected net utility has to be multiplied by (1−G(s)).
The equilibrium cutoff x∗ is such that UA(x∗) = UB(x∗). Solving for x∗
yields
x∗ =β
1− β[δEu− (1− δ)s] + δ(1−G(s)) (Eu[x | x > s]− s) +
11− β
s. (4)
Under the adjusted assumptions for case 2,
β = 1− 2− α
2− µand δ =
α
(2− µ)(1−G(s)).
Moreover, µ = 2 − 2G(x∗) must hold for x∗ to be an equilibrium. Replacing
these probabilities in (4) reveals that there will be no neat general solution.
7.3 Equilibrium with Uniform and Equal Capacities
So as to simplify further, we assume now that the two schools have equal ca-
pacities (i.e. we set α = 1), normalize M = 1 and let G be uniform on [0, 1].
Given the normalization M = 1, the requirement x∗(s) < M implies
s <4−√10
3= 0.279, (5)
and the cutoff is given as the relevant solution to the quadratic equation (4),
which is
x∗(s) ≡ 4s + 1− 2s2 +√
32s2 − 8s− 20s3 + 1 + 4s4
4(1− s). (6)
Also, G(x∗) = x∗, so that µ∗(s) = 2(1−G(x∗(s))) = 2(1− x∗(s)). Hence,
µ∗(s) =3 + 2s2 − 8s−√32s2 − 8s− 20s3 + 1 + 4s4
4(1− s). (7)
7 SWITCHING COSTS 34
Figure 5: The Number of Applicants to A.
7.4 Welfare
We first look at welfare under a static mechanism. Then we compute welfare
under the modified dynamic mechanism and compare them. Lastly, we derive
first-best welfare and compare welfare under the static mechanism and under
the dynamic mechanism with first-best.
Under the optimal static mechanism, individuals are randomly assigned to
a school for two periods. Expected utility of an individual in school A is thus
simply Eu = M2 = 1
2 per period. Over the two periods, students of mass 2 enjoy
this utility, so that
WSM = 1.
Under the dynamic mechanism, overall welfare consists of three components,
period one welfare denoted as W1, period two welfare denoted as W2 and ag-
gregate switching costs, which we denote by SWC. In general terms, period
one and two welfare are given as
W1 = µ∗E[x | x > x∗] + (1− µ∗)E[x | x ≤ x∗]
W2 = E[x | x ≥ s],
where the assumption that condition (1) holds, i.e.
(2− µ∗)(1−G(s)) ≥ 1
has been made implicitly. This assumption says that the number of individuals
from school B who do not opt out in t = 2 exceeds the capacity of school
A. Consequently, the period two allocation will assign only individuals from
7 SWITCHING COSTS 35
Figure 6: Aggregate Welfare W1 + W2.
school B to school A. Therefore, under this assumption, in period two students
of mass one move from B to A, and all students in A move to B, so that
aggregate switching costs are
SWC = 2s.
Now, G(.) being uniform and M = 1, we have G(s) = s,
E[x | x ≥ s] =1 + s
2, E[x | x > x∗] =
1 + x∗
2and E[x | x ≤ x∗] =
x∗
2,
while x∗ and µ∗ are given by equations (4) and (7) above; Figure 5 depicts
µ∗. It is straightforward to check that under these conditions, the assumption
(2− µ∗)(1−G(s)) ≥ 1 does indeed hold.
Computing W1 + W2 − 2s yields
14s2 − 28s + 11−√4s4 + 32s2 − 20s3 + 1− 8s
8(1− s).
Probably more conclusive is a look at Figures 6 and 7. Figure 6 depicts W1+W2.
The figure corroborates the conjecture made above that increasing s has also
a positive effect in that it makes the period two allocation more efficient if one
neglects the actual cost of switching schools. This is illustrated by the increasing
part of the curve W1 + W2 in Figure 6. However, the negative effect of a less
efficient selection due to a higher cutoff x∗(s) dominates when s becomes large.
Figure 7 depicts welfare under the modified dynamic mechanism, taking
the cost of switching schools into account. The straight line that parallels
7 SWITCHING COSTS 36
the horizontal axes is the aggregate welfare under the static mechanism. It
is particularly noteworthy that welfare under the dynamic mechanism exceeds
welfare under the static mechanism for all s up to approximately 0.127. In
other words, for switching costs no larger than approximately one fourth of the
expected utility generated by the better school does the dynamic mechanism
still outperform the static mechanism.
To complete the analysis, let us compute first-best welfare. In period one,
all individuals with x ≥ 12 go to school A. Consequently, period one welfare
under first-best is
WFB1 =
34.
In period two, the first-best allocation is slightly more complicated. Because
switching schools is costly, only those students who are in B in t = 1 should go
to A in t = 2 whose utility for A is greater than the utility of those in A whom
they replace by 2s or more. Otherwise, the cost of switching schools (which is
borne by both the student who moves to and the one who moves away from A)
would not be offset by the increase in utility. That is, there is a x such that
all students who have been in A in t = 1 and whose second period utility is
x ≤ x move to B and all those students who have been in B in t = 1 and whose
second period utility exceeds x + 2s move to A. Since A’s capacity is one and
since total mass of students both in A and B is one, the additional constraint
for the first-best scheme is
1−G(x) + 1−G(x + 2s) = 1 ⇔ G(x) + G(x + 2s) = 1.
Because G(.) is uniform,
x =12− s
is readily established. The total number of students who switch schools is given
by G(x) + 1 − G(x + 2s), which is equal to 1 − 2s. Thus, aggregate switching
costs are
SWCFB = s(1− 2s).
Clearly, this requires s ≤ 12 . For s > 1
2 , the first-best mechanism is static, i.e.
it induces no school changes in t = 2. Period two welfare is
WFB2 = (1−G(x))E[x | x ≥ x] + (1−G(x + 2s))E[x | x ≥ x + 2s].
Noting that (1 − G(x)) = 12 + s and (1 − G(x + 2s)) = 1
2 − s and replacing
E[x | x ≥ x] and E[x | x ≥ x+2s] by 34− s
2 and 34 + s
2 , one gets after simplifying
WFB2 =
34− s2.
7 SWITCHING COSTS 37
Figure 7: Aggregate Welfare W1 + W2 − SWC.
Overall, first-best welfare is thus
WFB ≡ WFB1 + WFB
2 − SWCFB =32− s + s2.
Let us compare first-best with welfare under the dynamic and under the static
mechanisms. We do this by plotting the welfare under the dynamic and the
static mechanism as fraction of the first-best welfare, i.e. by plotting W DM
W FB andW SM
W FB as functions of s in Figure 8. The parallel to the horizontal axes depicts
first-best.
For small switching costs, the dynamic mechanism achieves more than eighty
percent of the first-best welfare, while the static mechanism achieves less than
seventy percent of first-best. Notice also that for s = 0 the difference between
first-best and the dynamic mechanism is due solely to the second period since
in period one, the dynamic mechanism achieves first-best. As seen above, the
point of intersection of W DM
W FB and W SM
W FB is at s = 0.127, where the static and
dynamic mechanism achieve slightly more than seventy percent of first-best.
7.5 Other Equilibria
We now address the question whether there are other equilibria. We show that
there is no case 3 and no case 4 equilibrium with equal capacities (i.e. for α = 1)
and G(x) such that Eu = m.
No Case 3 and 4 Equilibrium To see that there is no case 3 and 4 equi-
librium when G is such that Eu = m, note first that for α = 1 the restric-
7 SWITCHING COSTS 38
Figure 8: Welfare Comparisons.
tions α ≤ min{F (x∗), 2 − F (x∗)} and α ≥ min{F (x∗), 2 − F (x∗)} require
F (x∗) = 1 ⇔ x∗ = m = Eu. Second, the cutoff in a case 3 and 4 equilib-
rium is given by
UA(x∗) = x∗+G(s)Eu−(1−G(s))s = (1−G(s))E[x | x > s]−s) = UB(x∗). (8)
This is easily seen when noting that because µ = 1, all applicants to A get into
A and all others get into B in period one. In period two, the fraction G(s) of
those who were in B in period one have a utility for A that is smaller than the
switching cost s. Consequently, they will opt out, so that their overall utility
is zero. With probability (1−G(s)) they have utility for A that outweighs the
cost of switching, which explains the expected utility of applying to B. Since
the fraction G(s) of applicants to B opts out, the probability of staying in A
for applicants to A is G(s), in which case their expected utility is Eu. With
the probability (1−G(s)), though, they will be assigned to B in period two, in
which case they bear the switching cost s. This explains the expression for the
expected utility of applying to A.
Solving (8) for x∗ yields
x∗ = (1−G(s))E[x | x > s]−G(s)Eu. (9)
So as to complete the proof that there is no case 3 and no case 4 equilibrium,
it suffices to show that x∗ < Eu. To see this, note that (1 − G(s))E[x |x > s] − G(s)Eu < Eu implies (1 − G(s))E[x | x > s] < Eu + G(s)Eu.
7 SWITCHING COSTS 39
But (1 − G(s))E[x | x > s] =∫ Ms xdG(x), which is certainly less than Eu +
G(s)Eu =∫ M0 xdG(x) + G(s)
∫ M0 xdG(x) for s > 0. Thus, x∗ < Eu follows.
For the uniform distribution on [0,1], for example, Eu = 12 , G(s) = s and
E[x | x > s] = 1+s2 . Thus, x∗ = 1−s−s2
2 , which is smaller than Eu = 12 for any
s > 0.
7.6 Modified Dynamic Mechanism
When looking at the allocation under the standard dynamic mechanism, there
is likely to be too much school switching in equilibrium, i.e. students switch
schools more often than seems necessary from an incentive point of view: Under
the case 2 equilibrium analyzed above, some students who apply to B in t = 1
are sent to A in t = 1 because of excess demand for B and sent back to B
in t = 2 because of excess demand for A by students with the same priority.
This is likely to be inefficient because the incentives to apply to B would not
necessarily be weakened if those rationed at B in t = 1 were assigned to A
permanently. Such a mechanism would induce less switching in equilibrium
and therefore may increase welfare. We are now going to show that such a
mechanism exists for a uniform distribution and equal capacities.
New Dynamic Mechanism The new mechanism works as follows. If you
apply to B in t = 1, you’ll have priority over A-applicants in t = 2. (This is as
before.) However, if you get rationed at the school you apply to in t = 1, you’ll
be in the school you are assigned to after being rationed for both periods.
Equilibrium As in a case 2 equilibrium, assume
(A1) µ < α < 2− µ ⇔ 2− µ > 2− α
and define the residual capacity of A as the capacity of A that remains to be
allocated after those who are in there permanently have been subtracted. Under
assumption (A1), the residual capacity of A is µ since α−µ seats are occupied
permanently by lucky B-applicants who were assigned to A in t = 1.
Note that (2−α)(1−G(s)) is the number of B-applicants who want to get
into A in t = 2 because for them, x > s holds. The second assumption is
(A2) µ ≤ (2− α)(1−G(s)).
This assumption implies that the number of B-applicants who want to get into
A in t = 2 exceeds the residual capacity of A. Consequently, no one who applied
to A in t = 1 will be in A in t = 2.
7 SWITCHING COSTS 40
Figure 9: Welfare under the New Dynamic Mechanism.
Under these assumptions, the cutoff equilibrium conditions are as follows.
Utility of applying to A is
UA(x) = x− s,
while utility of applying to B is
UB(x) =2− α
2− µ(1−G(s))
µ
(2− α)(1−G(s))(E[x | x > s]− s)
+(
1− 2− α
2− µ
)[x + Eu],
which simplifies to
UB(x) =µ
2− µ(E[x | x > s]− s) +
α− µ
2− µ[x + Eu].
Solving UA(x) = UB(x) yields
x∗ =µ(E[x | x > s]− 2s) + (α− µ)Eu + 2s
2− α.
Notice that for s = 0 we have x∗ = α2−αEu, which is the cutoff of a type 2
equilibrium.
For the uniform G(x) = x and α = 1, we get
x∗(s) =1− 2s
2(1− 3s)
7 SWITCHING COSTS 41
and
µ∗(s) =1− 4s
1− 3s.
Note that µ(s) > 0 requires s < 14 . For larger s, there is no interior cutoff, and
the mechanism reduces to a static mechanism where all students are randomly
and permanently assigned in t = 1.18 The assumption (A1) and (A2) are
satisfied because µ(s) < 1−s holds for all s > 0, which is (A2). Since for α = 1
(A1) is contained in (A2), (A1) is also satisfied. Thus, this is an interior cutoff
equilibrium for all s < 14 .
Welfare First note that aggregate switching costs are
SWC = 2µ(s)s = 2s1− 4s
1− 3s.
As conjectured, the aggregate switching costs under the new dynamic mecha-
nism are smaller than under the standard dynamic mechanism, which are 2s.
Second, welfare Wt in period t = 1, 2, neglecting switching costs, is
W1 = µE[x | x > x∗] + (1− µ)E[x | x < x∗]
W2 = (1− µ)Eu + µE[x | x > s].
These expressions are readily explained. The number of A-applicants assigned
to A in t = 1 is µ < 1. The expected utility of each of these applicants
is E[x | x > x∗], which explains the first term of W1. The second term is
expected utility of students who applied to B. Their expected utility for A is
E[x | x < x∗] and 1− µ of them are admitted to A, which gives us the second
term. As for W2, note first that there is no sorting among the 1 − µ students
with permanent seats. Thus, their expected utility is just the unconditional
expectation Eu, which gives us the first term. The residual capacity of µ is
filled with students from B whose utility exceeds s, which explains the second
term.
Substituting yields
W1 =3− 10s
4(1− 3s)
W2 =1− 2s− 4s2
2(1− 3s),
and aggregate welfare WDM,new, taking switching costs into account, is
WDM,new ≡ W1 + W2 − SWC =5 + 24s2 − 22s
4(1− 3s).
18To see this, note that with µ = 0, every one applies to B. Half of them get rationed andare assigned permanently to A.
8 SUMMARY 42
Assuming that a static mechanism would assign individuals for two periods,
overall welfare under a static mechanism is simply
WSM = 1.
Solving WDM,new = WSM yields (1− 6s)(1− 4s) = 0, which has two solutions16 and 1
4 . Thus, in the presence of switching costs the new dynamic mechanism
performs better than a static mechanism for all s < 16 . For s ∈ (
16 , 1
4
), the static
mechanism is better than the new dynamic mechanism, while for s ≥ 14 the two
mechanism induce the same equilibrium; see Figure 9.
8 Summary
For the various models analyzed in the paper, a number of results were obtained
that, coupled with the often subtle differences in the underlying assumptions,
may appear confusing to the reader. So as to make clear what results hold
under which conditions, we now provide a brief overview over the main results.
We broadly separate results into the categories Existence and Welfare.
8.1 Existence
Assume that all individuals agree that school A is better than school B in every
period. Proposition 1 guarantees the existence of an interior cutoff equilibrium
under fairly general conditions. That is, such an equilibrium exists for all
α < 2MM+Eu when utilities are uncorrelated over time, utility for school B is
normalized to zero and G has full support and satisfies G(0) = 0. Because G
has full support, Eu < M follows, implying 2MM+Eu > 1. Thus, Proposition 1
asserts in particular existence for all α ≤ 1. The same proposition states that
Eu = m is an alternative sufficient condition for equilibrium existence when
α > 2MM+Eu . An example that illustrates both the sufficiency and the relevance
of these conditions is in Appendix B. Proposition 12 in Appendix A shows
that for any α ∈ (0, 2) an interior cutoff equilibrium exists for any G satisfying
G(0) = 0 and full support when utility for B is not normalized but drawn
independently from G[0,M ] while utility for A is drawn from G[M, 2M ].
Under only somewhat more restrictive assumptions on α and ρ, the existence
result of Proposition 1 extends directly to the case when utilities are positively
correlated. This is the content of Proposition 8. For negative correlation,
Proposition 10 guarantees existence for any ρ and α, provided G is symmetric.
It may be worth mention at this point that symmetry of G guarantees existence
for all models (the only exception is the model with switching costs) independent
8 SUMMARY 43
of the degree or sign of correlation. Symmetry implies Eu = m. Consequently,
case 3 is an equilibrium for α ≤ 1 and case 4 is an equilibrium otherwise.
We have discussed uniqueness vs. multiplicity issues only for the model of
Section 3, where utilities are uncorrelated. The main result (Proposition 4) is
that under the dynamic mechanism the unique equilibrium for α ≤ 1 is a cutoff
equilibrium. For α > 1, there may be multiple equilibria with an interior cutoff,
or all students apply to B in period one.
The existence of a cutoff equilibrium when individuals differ with respect to
their ordinal preferences has been addressed in Section 4, where the assump-
tions were that some individuals have a time invariant preference for school A
and some for B and that utility of the disliked school is zero, while utility for
the preferred school is drawn i.i.d. from G[0,M ]. For this model, Lemma 2
shows that there is an appropriately adjusted dynamic mechanism such that all
individuals reveal their true ordinal preferences in phase 1 of this mechanism.
Assume that there is excess demand for one school after these ordinal prefer-
ences have been revealed (otherwise, the problem is trivially solved). Then,
Proposition 6 says that in phase 2 of the adjusted mechanism, the game re-
duces to the one where individuals of mass two have an ordinal preference for
the school with excess demand, which was studied in Section 3.
As for the model with switching costs, no general existence results were
obtained. Nonetheless, we showed for an example with uniform distribution
and equal capacities that a cutoff equilibrium exists when switching costs are
not too large.
8.2 Welfare
Assume that all individuals agree that school A is better than school B in every
period. Proposition 2, 9 and 11 address welfare under a dynamic and a static
mechanism for independent, positively and negatively correlated utility draws,
respectively. These propositions assert that in the interim stage of any cutoff
equilibrium, every individual has a greater expected utility under a dynamic
mechanism than in the equilibrium under a static mechanism. A corollary of
these propositions is that ex ante expected utility under the dynamic mechanism
is larger for any individual than under a static mechanism, provided, of course,
that there is a cutoff equilibrium.
When utility draws are independent over time, Proposition 3 says that when-
ever the allocation is first-best in period one under the dynamic mechanism,
9 CONCLUSIONS 44
it is second-best overall.19 This proposition relies heavily on independence of
utility draws over time because then achieving first-best in period one has no
opportunity cost in period two. This is not the case when utilities are corre-
lated, and consequently, we cannot make an equivalent statement for the case
with correlated utilities. For the model with i.i.d. utilities, Proposition 5 pro-
vides a lower bound for welfare that can be achieved in a T > 2 period model
by using dynamic mechanisms whenever a cutoff equilibrium exists in the two
period model.
Like with existence we haven’t obtained any general results with respect
to welfare when there are switching costs. For the example with the uniform
distribution and equal capacities, we showed that welfare under the dynamic
mechanism exceeds welfare under a static mechanism as long as switching costs
are moderate.
9 Conclusions
This paper studies the potential of dynamic mechanism to allocate indivisible
goods (e.g. houses or seats in a school) to individuals (e.g. students) when
these allocations are made repeatedly and when individuals face uncertainty
about the intensity of their future preferences.
For a two period two school model, where all individuals agree that one
school is better than the other, an equilibrium under the dynamic mechanism
exists under fairly general conditions. Moreover, we show that at the interim
stage every individual expects greater utility under the dynamic mechanism
than under a static mechanism, no matter what the degree of positive correla-
tion of first and second period utility. For the special case when both schools
have the same capacity and when the distribution that generates individuals’
instantaneous utilities is symmetric, equilibrium welfare is first-best in period
one and second-best over both periods.
In practice, a severe problem in school assignments with static mechanisms
is that bad schools are underdemanded since there are no incentives to apply to
these schools. This contrasts with assignments under the dynamic mechanism,
which provides exactly this type of incentives and thereby induces individuals to
apply to schools that are perceived as bad. Therefore, the dynamic mechanism
has also the potential of mitigating the problem that some schools have too
little demand.19First-best in period one is achieved, for example, when schools have equal capacities
(α = 1) and when G satisfies Eu = m.
9 CONCLUSIONS 45
At least two questions remain open and require further research. First,
though we derive conditions under which a dynamic mechanism establishes
the second-best allocation over both periods, little is known about the optimal
incentive compatible mechanism in our model when these conditions are not
met. This is a question we are currently working on. Second, the model with
more than two types of schools remains to be analyzed. An immediate extension
of the model with two schools and two periods is the following. Assume that
the number of schools N , each with a capacity of one, is equal to the number
of periods T and that the total mass of students is equal to N . Students
unanimously agree about the ordering of schools, which is A Â B Â ... Â N ,
and the cardinal utility xk for school k with k = A, ..,N is drawn i.i.d. over
time and schools, so that, e.g., xA ∼ G[(N − 1)M,NM ] and xN ∼ G[0,M ].
This model is balanced in an obvious sense.20 Consider the simple dynamic
mechanism: ”Each student can apply to every school exactly once.” That is,
after a student applied to A in t = 1 his choice set in t = 2 is {N, N − 1, ..., B},which is a direct extension of the dynamic mechanism of this paper to a balanced
N = T problem. There is an equilibrium under this mechanism which induces
no rationing at any stage in any school and in which welfare is first-best in
t = 1, better than under a random allocation in any t < T and equal to welfare
under a random allocation in t = T . In this sense, the results of the present
paper carry over to any balanced problem. Therefore, balanced models are a
natural starting point for the analysis of models with an arbitrary number of
schools and periods, possibly varying capacities and heterogenous preferences.
This analysis remains to be done.
20Appendix A contains this model with N = 2, except that it allows for varying capacities.
A THE MODEL WITHOUT THE NORMALIZATION 46
Appendix
A The Model without the Normalization
Throughout, we have maintained the assumption that utility for the worse
school is zero for all individuals and periods. We now relax this normalization
by making the following assumptions.
Total mass of students is two, and there are two periods and two schools A
and B. Capacity of A is α ∈ (0, 2) and capacity of B is (2− α). Instantaneous
utilities are i.i.d. draws from the distribution G(.) with support [0,M ] for B
and [M, 2M ] for A.21 As above, we denote by µ the mass of students who apply
to A in t = 1, and we denote now by ∆EU ≡ EUA − EUB > 0 the difference
between the two expected utilities. Note that ∆EU = M .
A.1 Equilibrium
The four cases:
As in the case with the normalization, as a function of α and µ four cases
are to be distinguished.
Case 1 : 2− µ < α < µ
Note that this implies 2− α < µ and 2− µ < 2− α. Then:
UA(x) =α
µxA +
(1− α
µ
)xB +
α− (2− µ)µ
EUA +(
1− α− (2− µ)µ
)EUB
= xB + EUA = UB(x).
Re-arranging and simplifying yields
x1A = xB +
2− α
α[EUA − EUB] = xB +
2− α
α∆EU.
Notice that the sole difference to case 1 with the normalization is that xB
appears on the right-hand side and that EUA is replaced by ∆EU .
Case 2 : µ < α < 2− µ
Note that this implies 2− µ > 2− α. Then:
UA(x) = xA + EUB
=2− α
2− µxB +
(1− 2− α
2− µ
)xA +
α
2− µEUA +
(1− α
2− µ
)EUB = UB(x).
Re-arranging and simplifying yields
x2A = xB +
α
2− α∆EU.
21Denote by GA the distribution of xA and by GB the distribution of xB . Then, ourassumption is that GA(M + x) = GB(x) for all x ∈ [0, M ].
A THE MODEL WITHOUT THE NORMALIZATION 47
Case 3 : α < min {µ, 2− µ}Note that this implies 2− µ < 2− α and µ < 2− α. Then:
UA(x) =α
µxA +
(1− α
µ
)xB + EUB
= xB +α
2− µEUA +
(1− α
2− µ
)EUB = UB(x).
Re-arranging and simplifying yields
x3A = xB +
µ
2− µ∆EU.
Case 4 : α > max {µ, 2− µ}Note that this implies 2− α < µ and 2− α < 2− µ. Then:
UA(x) = xA +2− α
µEUB +
(1− 2− α
µ
)EUA
=2− α
2− µxB +
(1− 2− α
2− µ
)xA + EUA = UB(x).
Re-arranging and simplifying yields
x4A = xB +
2− µ
µ∆EU.
Denote by x(φ) the set of all pairs (xA, xB) such that xA ≥ φ + xB for
φ ∈ [0, 2M ]. Formally,
x(φ) = {(xA, xB) | xA ≥ φ + xB, ∀xB ∈ [0,M ], ∀xA ∈ [M, 2M ]} .
There being a total mass of students equal to two,
F (x(φ)) = 2∫ M
0
∫ 2M
max{φ+xB ,M}dG(xA)dG(xB)
is the number of students with utility draws below (and to the the right of)
the line with slope 1 and the intercept φ on the horizontal axis. Because G(.)
is continuous in xB and xA, F (x(φ)) is also continuous for φ ∈ [0, 2M ]. In
particular, F (x(0)) = 0 and F (x(2M)) = 2. Moreover, since utility draws are
i.i.d., F (x(M)) = 1. Figure 10 illustrates the set x(φ) for φ = 0 and φ = M .
Replace µ by 2−F (xiA), where F (xi
A) is the mass of students with utilities
below the line given by xiA, i = 1, .., 4 and let
φ1 ≡ 2− α
α∆EU
φ2 ≡ α
2− α∆EU
φ3 ≡ 2− F (x(φ3))F (x(φ3))
∆EU
φ4 ≡ F (x(φ4))2− F (x(φ4))
∆EU.
A THE MODEL WITHOUT THE NORMALIZATION 48
Figure 10: The Model without the Normalization xB = 0.
Summarizing, we then have
x1A = xB +
2− α
α∆EU F (x(φ1)) ≤ α ≤ 2− F (x(φ1)
x2A = xB +
α
2− α∆EU 2− F (x(φ2)) ≤ α ≤ F (x(φ2))
x3A = xB +
2− F (x(φ3))F (x(φ3))
∆EU α ≤ min{F (x(φ3)), 2− F (x(φ3))}
x4A = xB +
F (x(φ4))2− F (x(φ4))
∆EU α ≥ max{F (x(φ4)), 2− F (x(φ4))}.
Proposition 12 For any α ∈ (0, 2) and any continuous G(.), there is an inte-
rior cutoff equilibrium.
Proof : Set φ = M . Then, F (x(φ)) = F (x(M)) = 1 and 2−F (x(M))F (x(M)) ∆EU =
M and F (x(M))2−F (x(M))∆EU = M . Thus, for α ≤ 1, case 3 is an equilibrium, and for
α > 1, case 4 is an equilibrium. ¥Remark Notice the difference to Proposition 1, which is valid for the model
with the normalization. Without the normalization, no additional restrictions
on α and G have to be made. This suggests that if the normalization does
anything it works against our mechanism.
A.2 Welfare with Equal Capacities
Let us now derive the first-best allocation, and show that this allocation is
the same in period one as the allocation under the dynamic mechanism when
B EXAMPLE OF EQUILIBRIUM NONEXISTENCE 49
schools have equal capacities. Thus, the result of the paper, according to which
first-best in period one is achieved exactly under these conditions, does not
appear to depend on the normalization we made.
First-best Consider an individual with utility draw x = (xA, xB). If this
individual is assigned to school A, it adds xA to total welfare. If assigned to B,
it adds xB, so that it adds net welfare of xA − xB when assigned to school A.
Maximizing aggregate welfare therefore requires assigning all those individuals
to A who add the largest net welfare when going to school A, subject to the
constraint that their mass does not exceed one, which is school A’s capacity. For
an illustration, consider Figure 10. Under first-best, all individuals with utility
draws above the line with φ = M are assigned to school A in both periods.
This is exactly what is achieved in the equilibrium under the dynamic mech-
anism in period one. All the individuals whose net welfare contributions are no
less than M go to school A, the other ones to school B. Thus, as in the case
with the normalization, first-best is achieved in t = 1, and second-best overall.
B Example of Equilibrium Nonexistence
This part of the Appendix provides an example that illustrates that the condi-
tions of Proposition 1 are sufficient and yet have grip.
Consider the piecewise uniform distribution
G(x) ={
θ2θ−1x if 0 ≤ x ≤ m
1− θ + θx if m < x ≤ 1,
with mean Eu = 3θ−14θ , median m = 2θ−1
2θ and the shape parameter θ > 12 .22
For θ = 1, we thus have Eu = m = 12 , which corresponds to the usual uniform
distribution. As θ approaches 12 , the median approaches zero and the mean 1
4 .
The difference between the expected value and the median is Eu−m = 1−θ4θ .
First, we demonstrate that an equilibrium may exist even when both suffi-
cient restrictions are violated. Let θ = 45 and α = 8
5 . Then, x2 = 74 > 1 ≡ M .
Thus, there is no case 2 equilibrium in this case. On the other hand, since θ 6= 1,
Eu 6= m, so the sufficient condition for the case 4 equilibrium is violated as well.
Nonetheless, there is a case 1 equilibrium with x1 = 764 . To see this, note that
x1 = 764 < 3
16 = m. Thus, F (x1) ≡ 2θ2θ−1x1 = 7
24 < 85 = α < 17
24 = 2 − F (x1).
Second, to see that the restrictions, though sufficient, have grip, assume now
θ = 35 and keep α = 8
5 . Then, x2 = 43 > 1 and x1 = 1
12 < 16 = m. Thus,
22It is easy to check that G(0) = 0 and G(m) = 12
holds.
B EXAMPLE OF EQUILIBRIUM NONEXISTENCE 50
F (x1) = 12 < 8
5 = α. However, 2 − F (x1) = 32 < 8
5 = α, so there is no case
1 equilibrium either. Though Eu 6= m, there might still be a case 4 equilib-
rium. However, no such equilibrium exists. For θ < 0.86, there is no solution
to x4 = F (x4)2−F (x4)Eu such that x4 > m. For x4 < m, the solution is easily seen
to be x4 = 5θ−33θ . But for θ = 3
5 , x4 = 0. Therefore, 2 − F (x4) = 2 > 85 = α,
violating the restrictions for a case 4 equilibrium. Since case 3 cannot be an
equilibrium when α > 1, it follows that there is no cutoff equilibrium.
REFERENCES 51
References
Abdulkadiroglu, A. (2004): “Better Mechanism Design and Implementa-
tion,” Working paper.
Abdulkadiroglu, A., P. Pathak, and A. Roth (2005): “The New York
City High School Match,” American Economic Review (Papers and Proceed-
ings), 95(2), 368–71.
Abdulkadiroglu, A., P. Pathak, A. Roth, and T. Sonmez (2005): “The
Boston Public School Match,” American Economic Review (Papers and Pro-
ceedings), 95(2), 364–7.
Abdulkadiroglu, A., and T. Sonmez (1998): “Random Serial Dictatorship
and the Core from Random Endowments in House Allocation Problems,”
Econometrica, 66(3), 689–701.
(1999): “House Allocation with Existing Tenants,” Journal of Eco-
nomic Theory, 88(2), 233–60.
(2003): “School Choice: A Mechanism Design Approach,” American
Economic Review, 93(3), 729–47.
Borgers, T., and P. Postl (2004): “Efficient Compromising,” Working pa-
per.
Boston Globe (2003): “School assignment flaws detailed,” December 12,
2003.
Casella, A. (2003): “Storable Votes,” Working paper.
Hatfield, J. W., and P. Milgrom (2005): “Auctions, Matching, and the
Law of Aggregate Demand,” American Economic Review, 95(4), 913–935.
Hortala-Vallve, R. (2004): “Qualitative Voting,” Working paper.
Jackson, M. O., and H. F. Sonnenschein (2004): “Overcoming Incentive
constraints by linking problems,” Working paper (extended version).
McLennan, A. (2002): “Ordinal Efficiency and the Polyhedral Separating
Hyperplane Theorem,” Journal of Economic Theory, 435-49.
Roth, A. E. (2002): “The Economist as Engineer: Game Theory, Experi-
mentation, and Computation as Tools for Design Economics,” Econometrica,
70(4), 1341–78.
REFERENCES 52
Roth, A. E., T. Sonmez, and M. U. Unver (2004): “Kidney Exchange,”
Quarterly Journal of Economics, 119(2), 457–88.
Roth, A. E., T. Sonmez, and M. U. Unver (2005a): “A Kidney Exchange
Clearinghouse in New England,” American Economic Review (Papers and
Proceedings), 95(2), 376–80.
Roth, A. E., T. Sonmez, and M. U. Unver (2005b): “Pairwise Kidney
Exchange,” Journal of Economic Theory (forthcoming).
Roth, A. E., and M. A. Sotomayor (1990): Two-Sided Matching: A
Study in Game Theoretic Modelling and Analysis. Econometric Society Mono-
graphs.
Sonmez, T., and M. U. Unver (2005): “Course Bidding at Business Schools,”
Boston College Working Paper.